Real LC::CalcCavityDist (size_t i, const std::string &name, const PropertySet &opts) {
Factor Bi;
Real maxdiff = 0;
if( props.verbose >= 2 )
cerr << "Initing cavity " << var(i) << "(" << delta(i).size() << " vars, " << delta(i).nrStates() << " states)" << endl;
if( props.cavity == Properties::CavityType::UNIFORM )
Bi = Factor(delta(i));
else {
InfAlg *cav = newInfAlg( name, *this, opts );
cav->makeCavity( i );
if( props.cavity == Properties::CavityType::FULL )
Bi = calcMarginal( *cav, cav->fg().delta(i), props.reinit );
else if( props.cavity == Properties::CavityType::PAIR ) {
vector<Factor> pairbeliefs = calcPairBeliefs( *cav, cav->fg().delta(i), props.reinit, false );
for( size_t ij = 0; ij < pairbeliefs.size(); ij++ )
Bi *= pairbeliefs[ij];
} else if( props.cavity == Properties::CavityType::PAIR2 ) {
vector<Factor> pairbeliefs = calcPairBeliefs( *cav, cav->fg().delta(i), props.reinit, true );
for( size_t ij = 0; ij < pairbeliefs.size(); ij++ )
Bi *= pairbeliefs[ij];
}
maxdiff = cav->maxDiff();
delete cav;
}
Bi.normalize();
_cavitydists[i] = Bi;
return maxdiff;
}
Real EMAlg::iterate( MaximizationStep &mstep ) {
Real logZ = 0;
Real likelihood = 0;
_estep.run();
logZ = _estep.logZ();
// Expectation calculation
for( Evidence::const_iterator e = _evidence.begin(); e != _evidence.end(); ++e ) {
InfAlg* clamped = _estep.clone();
// Apply evidence
for( Evidence::Observation::const_iterator i = e->begin(); i != e->end(); ++i )
clamped->clamp( clamped->fg().findVar(i->first), i->second );
clamped->init();
clamped->run();
likelihood += clamped->logZ() - logZ;
mstep.addExpectations( *clamped );
delete clamped;
}
// Maximization of parameters
mstep.maximize( _estep.fg() );
return likelihood;
}
/** Only maximization step included
*
* This function performs maximization based on the evidence collected from tab file
* Since there's no belief propagation stage, it does not return proper likelihood.
*/
Real UnsafeEmalg::iterateWithoutEstep(MaximizationStep &mstep) {
Real logZ = 0;
Real likelihood = 0;
// Expectation calculation
int nProcessed = 0;
for( Evidence::const_iterator e = _evidence.begin(); e != _evidence.end(); ++e ) {
InfAlg* clamped = _estep.clone();
// Apply evidence
for( Evidence::Observation::const_iterator i = e->begin(); i != e->end(); ++i )
clamped->clamp( clamped->fg().findVar(i->first), i->second );
mstep.addExpectations( *clamped );
delete clamped;
nProcessed++;
if (nProcessed % 10 == 0) {
cout << nProcessed << " number of samples processed" << endl;
}
}
// Maximization of parameters
mstep.maximize(_estep.fg());
return likelihood;
}
EMAlg::EMAlg( const Evidence &evidence, InfAlg &estep, std::istream &msteps_file )
: _evidence(evidence), _estep(estep), _msteps(), _iters(0), _lastLogZ(), _max_iters(MAX_ITERS_DEFAULT), _log_z_tol(LOG_Z_TOL_DEFAULT)
{
msteps_file.exceptions( std::istream::eofbit | std::istream::failbit | std::istream::badbit );
size_t num_msteps = -1;
msteps_file >> num_msteps;
_msteps.reserve(num_msteps);
for( size_t i = 0; i < num_msteps; ++i )
_msteps.push_back( MaximizationStep( msteps_file, estep.fg() ) );
}
int main() {
// This is the INFERENCE/EM engine, derived from the
// libDAI example program (example_sprinkler_em)
// (http://www.cs.ubc.ca/~murphyk/Bayes/bnintro.html)
//
// The factor graph file (input.fg) has to be generated first
// and the data file (sprinkler.tab),
// as well as the EM commands
// Read the factorgraph from the file
FactorGraph Network;
Network.ReadFromFile( "input.fg" );
// Prepare junction-tree object for doing exact inference for E-step
PropertySet infprops;
infprops.set( "verbose", (size_t)1 );
infprops.set( "updates", string("HUGIN") );
infprops.set( "maxiter", string("1000") );
infprops.set( "tol", string("0.00001") );
infprops.set( "logdomain", true);
infprops.set( "updates", string("SEQFIX") );
InfAlg* inf = newInfAlg("BP", Network, infprops );
inf->init();
// Read sample from file
Evidence e;
ifstream estream( "input.tab" );
e.addEvidenceTabFile( estream, Network );
cerr << "Number of samples: " << e.nrSamples() << endl;
// Read EM specification
ifstream emstream( "input.em" );
EMAlg em(e, *inf, emstream);
// Iterate EM until convergence
while( !em.hasSatisfiedTermConditions() ) {
Real l = em.iterate();
cerr << "Iteration " << em.Iterations() << " likelihood: " << l <<endl;
Real c = inf->logZ();
cerr << "Iteration infAlg " << em.Iterations() << " likelihood: " << c <<endl;
}
cout.precision(6);
cout << inf->fg();
delete inf;
return 0;
}
int main( int argc, char** argv ) {
if( argc != 4 )
usage("Incorrect number of arguments.");
FactorGraph fg;
fg.ReadFromFile( argv[1] );
PropertySet infprops;
infprops.set( "verbose", (size_t)0 );
infprops.set( "updates", string("HUGIN") );
InfAlg* inf = newInfAlg( "JTREE", fg, infprops );
inf->init();
Evidence e;
ifstream estream( argv[2] );
e.addEvidenceTabFile( estream, fg );
cout << "Number of samples: " << e.nrSamples() << endl;
for( Evidence::iterator ps = e.begin(); ps != e.end(); ps++ )
cout << "Sample #" << (ps - e.begin()) << " has " << ps->size() << " observations." << endl;
ifstream emstream( argv[3] );
EMAlg em(e, *inf, emstream);
while( !em.hasSatisfiedTermConditions() ) {
Real l = em.iterate();
cout << "Iteration " << em.Iterations() << " likelihood: " << l <<endl;
}
cout << endl << "Inferred Factor Graph:" << endl << "######################" << endl;
cout.precision(12);
cout << inf->fg();
delete inf;
return 0;
}
Real EM_estep(MaximizationStep &mstep, const Evidence &evidence, InfAlg &inf) {
Real likelihood = 0;
inf.run();
Real logZ = inf.logZ();
// Expectation calculation
for (Evidence::const_iterator e = evidence.begin(); e != evidence.end(); ++e) {
InfAlg* clamped = inf.clone();
// Apply evidence
for (Evidence::Observation::const_iterator i = e->begin(); i != e->end(); ++i)
clamped->clamp(clamped->fg().findVar(i->first), i->second);
clamped->init();
clamped->run();
likelihood += clamped->logZ() - logZ;
mstep.addExpectations(*clamped);
delete clamped;
}
return likelihood;
}
vector<Factor> allBeliefs() {
vector<Factor> result;
for( size_t i = 0; i < obj->fg().nrVars(); i++ )
result.push_back( obj->belief( obj->fg().var(i) ) );
return result;
}
Real BBPCostFunction::evaluate( const InfAlg &ia, const vector<size_t> *stateP ) const {
Real cf = 0.0;
const FactorGraph &fg = ia.fg();
switch( (size_t)(*this) ) {
case CFN_BETHE_ENT: // ignores state
cf = -ia.logZ();
break;
case CFN_VAR_ENT: // ignores state
for( size_t i = 0; i < fg.nrVars(); i++ )
cf += -ia.beliefV(i).entropy();
break;
case CFN_FACTOR_ENT: // ignores state
for( size_t I = 0; I < fg.nrFactors(); I++ )
cf += -ia.beliefF(I).entropy();
break;
case CFN_GIBBS_B:
case CFN_GIBBS_B2:
case CFN_GIBBS_EXP: {
DAI_ASSERT( stateP != NULL );
vector<size_t> state = *stateP;
DAI_ASSERT( state.size() == fg.nrVars() );
for( size_t i = 0; i < fg.nrVars(); i++ ) {
Real b = ia.beliefV(i)[state[i]];
switch( (size_t)(*this) ) {
case CFN_GIBBS_B:
cf += b;
break;
case CFN_GIBBS_B2:
cf += b * b / 2.0;
break;
case CFN_GIBBS_EXP:
cf += exp( b );
break;
default:
DAI_THROW(UNKNOWN_ENUM_VALUE);
}
}
break;
} case CFN_GIBBS_B_FACTOR:
case CFN_GIBBS_B2_FACTOR:
case CFN_GIBBS_EXP_FACTOR: {
DAI_ASSERT( stateP != NULL );
vector<size_t> state = *stateP;
DAI_ASSERT( state.size() == fg.nrVars() );
for( size_t I = 0; I < fg.nrFactors(); I++ ) {
size_t x_I = getFactorEntryForState( fg, I, state );
Real b = ia.beliefF(I)[x_I];
switch( (size_t)(*this) ) {
case CFN_GIBBS_B_FACTOR:
cf += b;
break;
case CFN_GIBBS_B2_FACTOR:
cf += b * b / 2.0;
break;
case CFN_GIBBS_EXP_FACTOR:
cf += exp( b );
break;
default:
DAI_THROW(UNKNOWN_ENUM_VALUE);
}
}
break;
} default:
DAI_THROWE(UNKNOWN_ENUM_VALUE, "Unknown cost function " + std::string(*this));
}
return cf;
}
/// Run the algorithm and store its results
void doDAI() {
double tic = toc();
if( obj != NULL ) {
// Initialize
obj->init();
// Run
obj->run();
// Record the time
time += toc() - tic;
// Store logarithm of the partition sum (if supported)
try {
logZ = obj->logZ();
has_logZ = true;
} catch( Exception &e ) {
if( e.getCode() == Exception::NOT_IMPLEMENTED )
has_logZ = false;
else
throw;
}
// Store maximum difference encountered in last iteration (if supported)
try {
maxdiff = obj->maxDiff();
has_maxdiff = true;
} catch( Exception &e ) {
if( e.getCode() == Exception::NOT_IMPLEMENTED )
has_maxdiff = false;
else
throw;
}
// Store number of iterations needed (if supported)
try {
iters = obj->Iterations();
has_iters = true;
} catch( Exception &e ) {
if( e.getCode() == Exception::NOT_IMPLEMENTED )
has_iters = false;
else
throw;
}
// Store variable marginals
varMarginals.clear();
for( size_t i = 0; i < obj->fg().nrVars(); i++ )
varMarginals.push_back( obj->beliefV( i ) );
// Store factor marginals
facMarginals.clear();
for( size_t I = 0; I < obj->fg().nrFactors(); I++ )
try {
facMarginals.push_back( obj->beliefF( I ) );
} catch( Exception &e ) {
if( e.getCode() == Exception::BELIEF_NOT_AVAILABLE )
facMarginals.push_back( Factor( obj->fg().factor(I).vars(), INFINITY ) );
else
throw;
}
// Store all marginals calculated by the method
allMarginals = obj->beliefs();
};
}
